OpenAI just wants to win

OpenAI said this week that an advanced, unreleased model it deployed produced a solution to the Navier-Stokes problem, one of the seven Clay Mathematics Institute Millennium Prize questions. The company told reporters it used a large fleet of agents and substantial compute over a short period to reach the result, positioning the achievement as a major technical milestone. The announcement immediately reignited tensions between the company and parts of the mathematics community. Several mathematicians interviewed by The Verge described growing unease with OpenAI’s rapid, resource-heavy incursions into difficult mathematical problems. At the center of the current dispute are claims about how OpenAI reacted when it learned independent researchers were making progress on the same problem. NYU’s Tristan Buckmaster says he and collaborator Levent Alpöge were pursuing Navier-Stokes when OpenAI became aware of their work. Buckmaster alleges that conversations with OpenAI researcher Sébastien Bubeck grew contentious and that the company offered him extensive compute and the chance to be sole author on OpenAI’s announcement — an offer Buckmaster says would have excluded Alpöge and that he viewed as a bribe. He also questioned whether his use of OpenAI’s Codex tool could have contributed to the company’s success; OpenAI has denied that Buckmaster’s recent Codex prompts could have influenced the system’s training and said no one accessed his specific user data, though some uncertainty about derivative uses of product data has been acknowledged in later comments. Bubeck disputes Buckmaster’s account of the conversations and says OpenAI made similar arrangements with other mathematicians, while noting that Alpöge’s employment at rival Anthropic made his involvement a problem from OpenAI’s perspective. Buckmaster declined the company’s offer, made his own work public, and said he assembled colleagues and lawyers to scrutinize both the mathematics and OpenAI’s conduct. The stakes are high: the Clay Institute established seven Millennium Prize problems in 2000, each carrying a $1 million award, and only one — the Poincaré conjecture — has been resolved since the list was published. For AI labs seeking to demonstrate mathematical superiority, a verified solution to a Millennium problem would be a major trophy, a fact that helps explain why researchers and companies alike rushed once rumors of progress began to circulate. The episode has left parts of the mathematical community wary of how powerful AI firms pursue breakthroughs and of the potential consequences for collaboration and established research norms.